Enhancing Landslide Dam Stability Prediction: A Data-Driven Framework Integrating Missing Data Imputation and Optimal Threshold Discrimination

Xiaojun Li¹Xiaobo Zhang²Jun He^3*Yi-Xiang Song⁴Yangqi Li⁵

• ¹Shanxi Vocational University of Engineering Science and Technology, Jinzhong, China
• ²China Academy of Building Research, Beijing, China
• ³Hebei University of Technology, Beichen District, China
• ⁴Hebei University of Technology, Tianjin, China
• ⁵Hebei Polytechnic of Building Materials, Qinhuangdao, China

Accurate prediction of the stability of landslide dams is crucial in preventing and mitigating potential threats to downstream communities. However, the development of reliable predictive models is hindered by incomplete landslide dam inventory datasets due to missing data. To overcome this challenge, this study proposes a datadriven approach that incorporates missing data imputation to enhance the applicability and accuracy of landslide dam stability predictions. On the basis of the collected landslide dam inventory containing 518 cases with a probability of missing rate of 25%, various data imputation methods including generative adversarial imputation Nets (GAIN), missForest, multiple imputations by chained equations (MICE), K-nearest neighbors (KNN) and mean most-frequency (MMF) were used to estimate missing values to improve the completeness of the datasets.The imputed datasets were used to predict the stability of landslide dams via various machine learning approaches (support vector machine (SVM), random forests (RF), extreme gradient boosting (XGBoost), and logistic regression (LR)). Our key innovation lies in coupling GAIN with SVM, enhanced by Youden-index-based threshold optimization for stability classification. Key results demonstrate GAIN's superiority: it achieved the lowest RMSE (0.205) for continuous variables and 66.0% accuracy for categorical data. The GAIN-SVM combination yielded the highest predictive performance (AUC = 0.823), outperforming traditional methods by 15.2%. The Youden-index further improved classification accuracy by 3.1-9.3% for ambiguous cases (probabilities ~0.5), addressing a critical gap in existing models. This framework enables rapid stability assessments even with incomplete field data, providing critical support for emergency decision-making in landslide-prone regions. It also allows reliable risk assessments in data-scarce regions, supporting timely hazard mitigation decisions.